Abstract:Multi-agent systems built on large language models have shown strong performance on complex reasoning tasks, yet most work focuses on agent roles and orchestration while treating inter-agent communication as a fixed interface. Latent communication through internal representations such as key-value caches offers a promising alternative to text-based protocols, but existing approaches do not jointly optimize communication with multi-agent reasoning. Therefore we propose DiffMAS, a training framework that treats latent communication as a learnable component of multi-agent systems. DiffMAS performs parameter-efficient supervised training over multi-agent latent trajectories, enabling agents to jointly learn how information should be encoded and interpreted across interactions. Experiments on mathematical reasoning, scientific QA, code generation, and commonsense benchmarks show that DiffMAS consistently improves reasoning accuracy and decoding stability over single-agent inference, text-based multi-agent systems, and prior latent communication methods, achieving 26.7% on AIME24, 20.2% on GPQA-Diamond, and consistent gains across reasoning benchmarks.




Abstract:Spurious correlations in training data significantly hinder the generalization capability of machine learning models when faced with distribution shifts in real-world scenarios. To tackle the problem, numerous debias approaches have been proposed and benchmarked on datasets intentionally designed with severe biases. However, it remains to be asked: \textit{1. Do existing benchmarks really capture biases in the real world? 2. Can existing debias methods handle biases in the real world?} To answer the questions, we revisit biased distributions in existing benchmarks and real-world datasets, and propose a fine-grained framework for analyzing dataset bias by disentangling it into the magnitude and prevalence of bias. We observe and theoretically demonstrate that existing benchmarks poorly represent real-world biases. We further introduce two novel biased distributions to bridge this gap, forming a nuanced evaluation framework for real-world debiasing. Building upon these results, we evaluate existing debias methods with our evaluation framework. Results show that existing methods are incapable of handling real-world biases. Through in-depth analysis, we propose a simple yet effective approach that can be easily applied to existing debias methods, named Debias in Destruction (DiD). Empirical results demonstrate the superiority of DiD, improving the performance of existing methods on all types of biases within the proposed evaluation framework.